ABSTRACT: Starvation, Latency, Overhead & Contention are the critical bottlenecks keeping us from achieving ExaScale computing; the ParalleX model of computation is being developed to address these challenges by enabling a new way of computation based on message-driven flow of distributed control in a global address space cooperated by lightweight synchronization elements. As an execution model, ParalleX is a set of governing principles that guide the co-design, function, and inter-interoperability of all layers of the system structure from the programming model through the system software to the hardware architecture. Key concepts of ParalleX are (1) an Active-Global-Address-Space, (2) parallel-processes, (3) multiple-ephemeral-threads, (4) Local-Control-Objects, (5) Parallel-Control-Elements, and (6) Percolation.
The current manifestation of ParalleX is a system-software implemented in C++ as a high-performance-parallex (HPX) to facilitate concepts-testing and performance-assessing of the synergy of PX-mechanisms.
The poster details ParalleX details and shows a preliminary-yet-significant scaling-test of ParalleX thread-management and a discusses future-work.